Model evaluation from a checkpoint with Multi GPU - tensorflow

Question

I know how to train a network on a single GPU -> save a checkpoint -> later on load this checkpoint -> run benchmarks.
I can't figure how to do it when I train using multiple GPUs and using the new Data API.
Here is the 'normal' training code:
import tensorflow as tf
images_placeholder = tf.placeholder(tf.float32, shape=(None, image_size,
image_size, 1), name='input')
labels_placeholder = tf.placeholder(tf.int32, shape=(None))
embeddings = build_graph(images_placeholder)
loss = add_loss(embeddings, labels_placeholder)
embeddings = tf.identity(embeddings, 'embeddings')
Later on, when I want to benchmark:
with tf.Graph().as_default():
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True)) as sess:
load_graph_def(model_path) # for example: d:\model.ckpt-0
images_placeholder = tf.get_default_graph().get_tensor_by_name("input:0")
embeddings = tf.get_default_graph().get_tensor_by_name("embeddings:0")
images = benchmark_utils.load_data(paths_batch, image_size)
feed_dict = {images_placeholder: images}
predictions = sess.run(embeddings, feed_dict=feed_dict)
So now I want to train with multiple GPUs like so:
with tf.Graph().as_default(), tf.device('/cpu:0'):
dataset = tf.data.Dataset.from_tensor_slices((images_list, labels_list))
dataset = dataset.map(load_images)
dataset = dataset.shuffle(buffer_size=100)
dataset = dataset.batch(128)
dataset = dataset.repeat()
opt = tf.train.MomentumOptimizer(0.01, momentum=0.9, use_nesterov=True)
tower_grads = []
with tf.variable_scope(tf.get_variable_scope()):
for i in range(num_gpus):
with tf.device('/gpu:%d' % i):
with tf.name_scope('%s_%d' % (TOWER_NAME, i)) as scope:
image_batch, label_batch = dataset.iterator.get_next()
loss = tower_loss(scope, image_batch, label_batch)
What I can't figure out is how can I get the 'input' and 'embeddings' tensor when I want to benchmark the checkpoint.
How do I define for example the tensor called 'input' that should receive the images that should be evaluated ?
I'm guessing that somewhere in the multi-gpu code, I should define this images_placeholder like I defined in the single-gpu training.
Thanks for any advice!